Artificial intelligences detect adverse drug effects in medical records

Electronic medical records contain a wealth of information about adverse drug effects that often elude pharmacovigilance databases. A recent study demonstrated that an advanced artificial intelligence model can analyze these records to identify links between a drug and a side effect with remarkable accuracy.

The tool was tested on 372 varied medical reports, from different specialties and clinical contexts. Of the 191 adverse effects identified by the model, 180 were accurate, achieving a precision rate of 94.2%. Additionally, the system correctly identified the severity of the effects in 100% of cases and their undocumented nature in 93.9% of situations. Approximately 12% of the detected effects were considered serious according to the Food and Drug Administration criteria, and 15% were not mentioned in the drug leaflets.

The cost of this analysis remains moderate: on average, $0.18 per record and $0.35 per validated adverse effect. The model even detected effects not reported by physicians, such as a magnesium drop linked to the use of tacrolimus, an immunosuppressive drug.

Unlike traditional methods, which rely on fixed rules or sentence-by-sentence analysis, this approach uses a model capable of understanding the overall context of a report. It takes into account subtle clues, such as physicians’ decisions to stop or adjust treatment due to an adverse effect. This makes it possible to distinguish true cause-and-effect relationships from mere coincidences.

The results show that this technology could transform post-market drug monitoring. By integrating such tools into existing systems, such as those used by health authorities, it would be possible to detect safety signals more quickly, including rare or serious undocumented effects. This would pave the way for better patient protection against treatment-related risks.

The model also demonstrated its ability to classify adverse effects according to their severity, unexpected nature, or connection to the patient’s current illness. For example, it correctly identified that some effects, although not mentioned in the leaflets, were already known in the scientific literature. In other cases, it detected effects not yet associated with a drug, providing new avenues for research.

A challenge remains: the model can sometimes fabricate information, such as drug names or medical terms, especially when data is partially masked to protect privacy. However, these errors remain minor and do not significantly affect the overall quality of the results.

This advancement could therefore enable health authorities and pharmaceutical companies to better monitor drug safety in real time, by leveraging a previously underutilized source of information: the detailed clinical notes of healthcare professionals.

Sources and Credits

Source Study

DOI: https://doi.org/10.1007/s40264-026-01682-6

Title: A Large Language Model for Extracting Post-marketing Adverse Drug Events from Clinical Notes in the Electronic Health Record

Journal: Drug Safety

Publisher: Springer Science and Business Media LLC

Authors: Dana Ludwig; Michelle Wang; James Buchanan; Trang Trinh